T. Chirakitsakul et al.: Integration of Convolutional Neural Network and Image Processing
1) Eliminating False Fibrils Parallel to Fiber and Unwanted-Object Edges: Vertical and horizontal false fibrils adjacent to fiber or fine edges are removed using a sliding window technique. For each fibril pixel, a 1 × 7 window is applied to detect vertical false fibrils, and a 7 × 1 window is used for horizontal false fibrils. If any part of the window overlaps with fiber or fine pixels, the fibril pixel in the center of the window is reclassified as a background pixel. 2) Eliminating False Fibrils Within Gaps of Fibers and Unwanted-Objects Using a CNN Model: Gaps within fibers and fines are areas prone to dense false fibrils because the translucent texture of fibers and fines closely resembles the appearance of fibrils. This similarity makes it challenging for traditional image processing techniques to determine whether the detected pixels are false fibrils based solely on local spatial information. To address this issue, a CNN model, YOLOv4, is employed to identify these gaps in the image. Fibril pixels located within the detected areas are excluded from the final results, as illus- trated in Figure 5. The YOLOv4 model is fine-tuned on a training dataset of 15 images with manually labeled bounding boxes. These 15 training images are augmented with three flipping transformations— left-to-right flip (horizontal flip), top-to-bottom flip (vertical flip), and a combined left-to-right and top-to- bottom flip (180-degree rotation)—resulting in a total of 60 training images. Despite achieving a detection rate of 12%, the model significantly reduces the false fibril rate from 8.44% to 1.76%. In the test images, most potential gaps with dense false fibrils are successfully detected. One possible reason for the low detection rate could be the over-labeling of the ground truth. 3) Removing Detached Fibrils Floating in the Back- ground: Floating fibrils that are not connected to fibers are excluded from the fibril detected result, as they do not contribute to the fibrillation index. This is achieved by analyzing pixel connectivity within a 13 × 13sliding window. Fibril pixels without neighboring fibril pixels within the window are reclassified as background. The sizeof13 × 13 window ensures that true fibril fragments are retained during this process. D. FIBRILLATION INDEX COMPUTATION Finally, the fibrillation index is calculated as the ratio of fibril area to fiber area, a reliable measure of the degree of external fibrillation of fibers. Such quantification is essential for understanding the effects of refining on fiber properties, as highlighted in various studies, including those investigating new measurement technologies for fiber quality control. As fibril area representing quantity of microfibrillar fines, the fibrillation index computation can be calculated as the percentage ratio of the number of fibril pixels to the
FIGURE 4. The result of fibrils detection process containing some false fibrils.
indicates a homogenous region, such as a clear background. As illustrated in Figure 3(a), the surface plot for a patch with a homogeneous background displays a relatively flat and uniform surface, indicating low variability in intensity differences. In contrast, the surface plot for a patch containing fibrils or noise exhibits greater variability, with peaks in the intensity difference values. This non-uniform pattern arises from the presence of fibrils or noise structures that disrupt the uniformity of pixel intensities, resulting in higher SD differences . Forty fibril patches and forty background patches selected from the training images, were used to determine the optimal threshold value, referred to as the predefined threshold T SD . This threshold is applied to identify and exclude clear background patches. Patches with SD differences values lower than T SD are classified as clear background patches, while those with values equal to or greater than T SD are classified as either fibril patches or noisy background patches. Experimental results indicated that setting T SD to 1.2 yielded the best classification performance. Following this step, the ResNet-50 is employed to separate patches containing fibrils from noisy patches without fibrils. CNNs excel in feature extraction and discrimination of local patterns, effectively addressing the challenges posed by the visual similarity between fibrils and noises. The ResNet-50 model used in this step is fine-tuned using the training dataset of 2,600 image patches, conprising 1,300 noisy background patches (class 0) and 1,300 fibril patches on both clear and noisy background (class 1), with an average accuracy of 96.30%. STEP3: Thresholding-based fibril segmentation Fibril detection is performed by applying Bradley’s adaptive thresholding [35] to identify fibril pixels in fibril patches. The detection results from all patches are then combined to produce a unified fibril detection map, as shown in Figure 4. C. POSTPROCESSING Postprocessing refines fibril detection results, focusing on eliminating false positives, ensuring accurate computation of the fibrillation index. We refine the fibril detection results through three consecutive steps as follows.
74640
VOLUME 13, 2025
Made with FlippingBook. PDF to flipbook with ease